readme_sdgnn.md

SDGNN: Learning Node Representation for Signed Directed Networks

Abstract

In this paper, we propose a novel Signed Directed Graph Neural Networks model named SDGNN to learn node embeddings for signed directed networks. The proposed model simultaneously reconstructs link signs, link directions, and signed directed triangles.

Overview

This paper is accepted at AAAI2021.

We provide a Pytorch implementation of SDGNN model. You can change different aggregator and other parameters to explore different architectures.

Comparison of Different Model Architectures

METHOD	SDGNN	BESIDES	SiGAT	SGCN
GNN Aggregator	Y	N	Y	Y
Layer By Layer	Y	N	N	Y
Direction Loss Function	Y	Y	N	N
Triangle Loss Function	Y	Y	N	N

Run code

Run python sdgnn.py .py to get results.

pos_ratio: 0.9394377842083506
accuracy: 0.9470855725506407
f1_score: 0.9720950512317418
macro f1_score: 0.731067445934596
micro f1_score: 0.9470855725506407
auc score: 0.8810108997247847

Some Notes

This code relies on Pytorch (>=1.0.1) due to the torch.spmm api

Bibtex

@inproceedings{huangsdgnn,
  title={SDGNN: Learning Node Representation for Signed Directed Networks},
  author={Huang, Junjie and Shen, Huawei and Hou, Liang and Cheng, Xueqi},
  booktitle={AAAI},
  year={2021}
}